Puppet-Master: Scaling Interactive Video Generation as a Motion Prior for Part-Level Dynamics
Paper
• 2408.04631 • Published
• 9
image imagewidth (px) 512 512 | label class label 100 classes |
|---|---|
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
000000 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
100001 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
200002 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
300003 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
400004 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
500005 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
600006 | |
700007 | |
700007 |
This test set is used to evaluate DragAPart and Puppet-Master.
Frame folders & Ground‑truth videos
Under the dataset root, each sample has its own sub‑folder.
Inside each sub‑folder you will find 14 PNG frames extracted from the ground‑truth video:
{sample_id}/
cXXX_tYYY.png
cXXX_tYYY.png
…
where cXXX is the zero‑padded camera index and tYYY is the zero‑padded frame index.
Metadata
A metadata.json file in the dataset root maps each sample ID to two fields:
reverse — true if the original motion runs backward (larger frame index first), otherwise false.per_frame_drags — the sparse motion trajectories as drags. It's an array of shape (14, 5, 4), where it specifies the starting coordinates and ending coordinates of the 5 drags on the 14 frames. The coordinates are normalized to [0, 1] range.Drag-a-Move-test-split/
├── 00000/
│ ├── c031_t000.png
│ ├── c031_t010.png
│ └── … (14 frames total)
├── 00001/
│ ├── c005_t020.png
│ └── …
├── …
├── metadata.json
└── README.md
@article{li2024dragapart,
title = {DragAPart: Learning a Part-Level Motion Prior for Articulated Objects},
author = {Li, Ruining and Zheng, Chuanxia and Rupprecht, Christian and Vedaldi, Andrea},
journal = {arXiv preprint arXiv:2403.15382},
year = {2024}
}
@article{li2024puppetmaster,
title = {Puppet-Master: Scaling Interactive Video Generation as a Motion Prior for Part-Level Dynamics},
author = {Li, Ruining and Zheng, Chuanxia and Rupprecht, Christian and Vedaldi, Andrea},
journal = {arXiv preprint arXiv:2408.04631},
year = {2024}
}